Rule extraction for multi bottom-up tree transducers

Following the invention of computers, it was always a dream to obtain translations automatically. If we give a machine a sentence it should return a sentence in another language expressing the same meaning. In the subfield of statistical machine translation (SMT), this translation is achieved with the help of statistical models. Those models use large text collections to automatically learn basic translation units that model the translation from a source sentence into a target sentence. The basic translation units can be single words or phrases consisting of multiple words. Other approaches, called syntax-based SMT, use rules of some formal grammar as their basic translation units. Syntax-based SMT systems easily allow the use of linguistic annotations. Rules can contain nonterminal symbols which can encode linguistic annotations. Furthermore, one can decide whether such annotations are used for both the source and target language, for one language only, or if those annotations are excluded alltogether. The integration of linguistic annotations yielded mixed results. In some cases translation quality significantly improves whereas in others it seems to hurt coverage and thus overall translation quality. While the use of annotations for both languages generally did not result in good translation quality, the use for one language only showed improvements. However, the best results are often obtained by a syntax-based SMT system that excludes all linguistic annotations. The underlying formal grammars vastly vary with respect to their expressive power. Synchronous context-free grammars are widely used but more powerful for-